English is currently available.
Raspberry Pi DISCUSSION
A camera application continues displaying a reasonable frame rate, but detections gradually lag behind live events. How can buffering be distinguished from a slow neural-network call, and which timestamp should the display use?
Reference: Raspberry Pi: AI software (https://www.raspberrypi.com/documentation/computers/ai.html).
Record capture time and result time for each frame. Throughput measures how many frames complete per second; latency measures how old each result is. An unbounded queue can grow while the application still appears busy and reports an acceptable processing rate.
For a live preview, consider a bounded latest-frame policy when processing cannot keep up, while keeping any recording requirements separate. Count dropped frames and preserve sequence numbers. Reducing queue length changes the loss-latency tradeoff, so it should be an explicit application decision.
Profile capture, conversion, inference and display with the supported Pi camera and accelerator stack. Test sustained operation with actual image dimensions and logging enabled. Show or log observation age so a stale detection is not presented as a current state.
